国产算力平台部署问题反馈

[myhloli]

我们新增了一些国产算力平台的支持，包括以下平台：

• 昇腾 Ascend
• 平头哥 T-Head
• 沐曦 METAX
• 海光 Hygon
• 燧原 Enflame
• 摩尔线程 MooreThreads
• 天数智芯 IluvatarCorex
• 寒武纪 Cambricon
• 昆仑芯 Kunlunxin
• 太初元碁 Tecorigin
• 壁仞 Biren

您可点击链接跳转至文档页查阅各平台的详细部署方案

国产硬件适配是一项复杂且具有挑战性的工作。我们已尽最大努力确保兼容性、功能完整性与运行稳定性，但仍可能存在个别场景下的稳定性问题、兼容性差异或精度对齐偏差。建议您在使用前，参考文档页面中的「红绿灯」状态标识，结合自身需求选择合适的平台与使用场景。

如在使用过程中遇到文档未覆盖的问题，欢迎在本帖中反馈——这将帮助更多用户快速定位和解决问题。
📌 发帖前请注意：
请先浏览已有回复，确认是否已有相同问题的讨论；
若为新问题，请单独开一个楼层发布，便于后续追踪与查阅。

感谢您的理解与支持！

[Xiaochaomeng]

大佬想请教下命令行的npu中vllm框架的红灯是体现在精度上吗，我在用旧版本的mineru后端vlm-vllm-engine调用do_parse已经适配好了会对精度有影响嘛

[Yikun]

@Xiaochaomeng 方便的话，提个issue以及复现用例，我们看下。争取能在0.11.0版本解决啊。

cc @wangxiyuan @gcanlin @shen-shanshan

[myhloli]

@Xiaochaomeng 方便的话，提个issue以及复现用例，我们看下。争取能在0.11.0版本解决啊。

cc @wangxiyuan @gcanlin @shen-shanshan

vllm-project/vllm-ascend#4502

[imwhyyyy]

好的,大佬我现在是修改部分vllm源码使用同步engine已经可以大批量解析了，请问会有什么隐患嘛
请问可以提供下您的源码修改吗

[myhloli]

@Yikun 目前通过预配显存参数跳过了命令行中的显存检测过程，可以在命令行中使用vllm的同步engine了。目前还有一个问题就是在fast-api中，如果先调用pipeline backend解析一本pdf后再调用vllm-engine，又会出现engine初始化失败的情况，目前解决方案是先用vllm-engine解析一本pdf，完成engine初始化后，再切换pipeline就没问题了。这种方法可以凑合着用，但不是很优雅，还是希望贵团队能够修复此问题。🙏

[myhloli]

环境变量指定spwan模式有效，已更新文档

[myhloli]

默认的基础镜像对应的是A2服务器910b卡的，300i可以尝试将基础镜像的tag更换为v0.11.0rc2-310pv0.10.0rc1-310p重新build再试一下

[trappc406-coder]

可以了，感谢 @myhloli

[bandiangou001]

怎么实现Mineru使用多张910B的npu显卡跑啊

[myhloli]

可以使用数据并行功能启动服务

[rade-f]

不好意思，请问910A设备有没有可以尝试的路子可以运行？

[myhloli]

不支持910a

[rade-f]

感谢回复

[zx1259]

请问 可以支持 310P3 的设备进行运行嘛

[myhloli]

基础镜像tag换成v0.11.0rc2-310pv0.10.0rc1-310p即可

[myhloli]

可以参考https://docs.vllm.ai/projects/ascend/en/latest/tutorials/single_node_300i.html

[myhloli]

可以参考https://docs.vllm.ai/projects/ascend/en/latest/tutorials/single_node_300i.html

[liuzeyu4201]

谢谢，尝试了文档的方法。
我尝试了直接拉起vllm server
命令如下：
root@user:/workspace# python -m vllm.entrypoints.openai.api_server
--model /root/.cache/modelscope/hub/models/OpenDataLab/MinerU2.5-2509-1.2B
--tensor-parallel-size 1
--enforce-eager
--dtype half
--block-size 128
--max-model-len 4096
--gpu-memory-utilization 0.3
--host 0.0.0.0
--port 9237
--disable-chunked-prefill
--disable-prefix-caching

不走mineru 也会出现同样的错误如下：
(EngineCore_DP0 pid=33771) WARNING 12-01 07:41:29 [platform.py:275] If chunked prefill or prefix caching is enabled, block size must be set to 128.
Loading safetensors checkpoint shards: 0% Completed | 0/1 [00:00<?, ?it/s]
Loading safetensors checkpoint shards: 100% Completed | 1/1 [00:01<00:00, 1.55s/it]
Loading safetensors checkpoint shards: 100% Completed | 1/1 [00:01<00:00, 1.55s/it]
(EngineCore_DP0 pid=33771)
(EngineCore_DP0 pid=33771) INFO 12-01 07:41:31 [default_loader.py:267] Loading weights took 1.73 seconds
(EngineCore_DP0 pid=33771) INFO 12-01 07:41:32 [model_runner_v1.py:2667] Loading model weights took 2.1678 GB
(EngineCore_DP0 pid=33771) WARNING 12-01 07:41:34 [cudagraph_dispatcher.py:106] cudagraph dispatching keys are not initialized. No cudagraph will be used.
(EngineCore_DP0 pid=33771) INFO 12-01 07:41:35 [worker_v1.py:256] Available memory: 9547198873, total memory: 46381125632
(EngineCore_DP0 pid=33771) INFO 12-01 07:41:35 [kv_cache_utils.py:1087] GPU KV cache size: 776,832 tokens
(EngineCore_DP0 pid=33771) INFO 12-01 07:41:35 [kv_cache_utils.py:1091] Maximum concurrency for 4,096 tokens per request: 189.66x
[rank0]:[W1201 07:41:36.262587355 compiler_depend.ts:62] Warning: Cannot create tensor with NZ format while dim < 2, tensor will be created with ND format. (function operator())
[rank0]:[W1201 07:41:36.573299850 compiler_depend.ts:57] Warning: EH9999: Inner Error!
The error from device(0), serial number is 16. there is a sdma error, sdma channel is 0, the channel exist the following problems: The SMMU returns a Terminate error during page table translation.. the value of CQE status is 2. the description of CQE status: When the SQE translates a page table, the SMMU returns a Terminate error.it's config include: setting1=0xc000000880e0000, setting2=0xff009000ff004c, setting3=0, sq base addr=0x800d00001003d000[FUNC:ProcessSdmaErrorInfo][FILE:device_error_proc.cc][LINE:780]
rtStreamSynchronize execute failed, reason=[sdma copy error][FUNC:FuncErrorReason][FILE:error_message_manage.cc][LINE:53]
EH9999[PID: 33771] 2025-12-01-07:41:36.369.861 (EH9999): synchronize stream failed, runtime result = 507013[FUNC:ReportCallError][FILE:log_inner.cpp][LINE:162]
TraceBack (most recent call last):

DEVICE[0] PID[33771]:
EXCEPTION STREAM:
Exception info:TGID=2652498, model id=65535, stream id=3, stream phase=SCHEDULE
Message info[0]:RTS_HWTS: hwts sdma error, slot_id=3, stream_id=3
Other info[0]:time=2025-12-01-15:41:35.667.162, function=int_process_hwts_sdma_error, line=2070, error code=0x20b (function copy_between_host_and_device_opapi)

[liuzeyu4201]

问题解决了：vllm-ascend:v0.11.0rc2-310p对300Iduo不支持。降级到quay.io/ascend/vllm-ascend:v0.10.0rc1-310p就行了
vllm-project/vllm-ascend#4026

[myhloli]

310p需要在命令行中添加如下两个参数--enforce-eager --dtype float16

[dutianyu]

310p需要在命令行中添加如下两个参数--enforce-eager --dtype float16

及时雨👍️感谢感谢，以下命令的确可行

mineru-gradio --server-name 0.0.0.0 --server-port 9191 --enable-vllm-engine true --dtype float16 --enforce-eager

[Juruobudong]

请问下Atlas 300V Pro有适配吗，有什么启动参数需要注意的

[myhloli]

没有明确支持该型号，已经明确支持的型号是300i duo，您可以参考文档中300i duo的方式尝试部署 https://opendatalab.github.io/MinerU/zh/usage/acceleration_cards/Ascend/

[Juruobudong]

感谢回复@myhloli
请问下是参考这条嘛Atlas 300I Duo适配的版本,请将第三行的tag修改为 v0.10.0rc1-310p

[zerobing]

请问支持 海光 K100 么

[zerobing]

@myhloli 麻烦帮确认下

[myhloli]

目前没有适配

[myhloli]

据vllm-ascend官方文档说明，300i为实验性支持，具体表现可能与其他npu卡（如910b）等不一致，您可以在vllm-ascend项目获得更多信息与帮助。

[ymw123]

vlm-transformers支持吗，使用vllm双卡没有成功，想使用双卡，后端换成vlm-transformers，解析时还是有BFloat16的问题，已经在命令中添加了--dtype float16的。

root@ununtu:/workspace# mineru-gradio --server-name 0.0.0.0 --server-port 7860 --enforce-eager
--dtype float16 --tensor-parallel-size 1 --data-parallel-size 2 --gpu-memory-utilization 0.
6

onnxruntime cpuid_info warning: Unknown CPU vendor. cpuinfo_vendor value: 15

• Running on local URL: http://0.0.0.0:7860
• To create a public link, set share=True in launch().
  2025-12-22 12:14:46.907 | WARNING | mineru.utils.pdf_page_id:get_end_page_id:8 - end_page_id is out of range, use images length
  2025-12-22 12:15:09.287 | INFO | mineru.backend.vlm.utils:set_default_batch_size:92 - gpu_memory: 43 GB, batch_size: 8
  2025-12-22 12:15:09.289 | INFO | mineru.backend.vlm.vlm_analyze:get_model:189 - get transformers predictor cost: 22.32s
  Layout Preparation: 100%|████████████████████████████████████████| 85/85 [00:00<00:00, 333.67it/s]
  Predict: 0%| | 0/85 [00:00<?, ?it/s].[W1222 12:15:19.609799468 compiler_depend.ts:47] Warning: The oprator of ne is executed, Currently High Accuracy but Low Performance OP with 64-bit has been used, Please Do Some Cast at Python Functions with 32-bit for Better Performance! (function operator())
  .[W1222 12:15:29.610186043 compiler_depend.ts:81] Warning: [Check][offset] Check input storage_offset[%ld] = 0 failed, result is untrustworthy239898084 (function operator())
  Predict: 0%| | 0/85 [00:15<?, ?it/s]
  [W1222 12:15:33.042176971 compiler_depend.ts:164] Warning: Device do not support double dtype now, dtype cast replace with float. (function operator())
  ./usr/local/python3.11.13/lib/python3.11/site-packages/torch/_tensor_str.py:397: UserWarning: current tensor is running as_strided, don't perform inplace operations on the returned value. If you encounter this warning and have precision issues, you can try torch.npu.config.allow_internal_format = False to resolve precision issues. (Triggered internally at build/CMakeFiles/torch_npu.dir/compiler_depend.ts:128.)
  start = [self[i] for i in range(0, PRINT_OPTS.edgeitems)]
  2025-12-22 12:15:30.150 | ERROR | mineru.cli.gradio_app:parse_pdf:50 - expected dtype Float but got dtype BFloat16
  [ERROR] 2025-12-22-12:15:30 (PID:37968, Device:0, RankID:-1) ERR01002 OPS invalid type

[myhloli]

transformes不支持vllm的命令行参数，也不支持数据并行，在vllm可用的情况下不推荐使用transformers推理

[ymw123]

文件上传已经快10分钟了，还在Added request这一步，这是正常的吗？为什么这么慢？[DeepWiki]搜索的时候提示使用vllm的时候不用理会“Using a slow image processor as use_fast is unset”，是这样吗？

2025-12-22 15:03:01.650 | WARNING | mineru.utils.pdf_page_id:get_end_page_id:8 - end_page_id is out of range, use images length
Two Step Extraction: 0%| | 0/439 [00:00<?, ?it/s]Using a slow image processor as use_fast is unset and a slow processor was saved with this model. use_fast=True will be the default behavior in v4.52, even if the model was saved with a slow processor. This will result in minor differences in outputs. You'll still be able to use a slow processor with use_fast=False.
INFO 12-22 15:03:32 [async_llm.py:269] Added request e43dc054-1121-4a84-86c8-477d191a95e6.
('Warning: torch.save with "_use_new_zipfile_serialization = False" is not recommended for npu tensor, which may bring unexpected errors and hopefully set "_use_new_zipfile_serialization = True"', 'if it is necessary to use this, please convert the npu tensor to cpu tensor for saving')
INFO 12-22 15:03:32 [async_llm.py:269] Added request 0dc1af3e-bb28-4407-8d59-7a05a337b7a1.
INFO 12-22 15:03:32 [async_llm.py:269] Added request e6f07548-e929-4938-9dd7-8583581a532e.
..INFO 12-22 15:03:32 [async_llm.py:269] Added request 9686c51d-3036-47c8-8603-731b10c29116.
INFO 12-22 15:03:33 [async_llm.py:269] Added request 69a338e5-42ee-4b6c-89fc-a8662693d49d.
INFO 12-22 15:03:33 [async_llm.py:269] Added request 1578b02b-57b8-44da-be27-68bc0b2a2764.
INFO 12-22 15:03:33 [async_llm.py:269] Added request 1061fd01-4672-49ea-a63a-3fcf49abae39.
INFO 12-22 15:03:33 [async_llm.py:269] Added request 2f83d898-64ad-4f0c-8b8b-27b5bc8c9634.

[pengxin99]

请问在昇腾平台上有什么性能优化吗？ @Yikun

[Yikun]

https://opendatalab.github.io/MinerU/zh/usage/acceleration_cards/Ascend/

按这个教程来，其中dockerfile中的版本，可以更新为FROM quay.m.daocloud.io/ascend/vllm-ascend:v0.11.0

[licyclone]

ppu的镜像构建的下载不了。 是什么原因？FROM crpi-vofi3w62lkohhxsp.cn-shanghai.personal.cr.aliyuncs.com/opendatalab-mineru/ppu:ppu-pytorch2.6.0-ubuntu24.04-cuda12.6-vllm0.8.5-py312
异常日志：msg="Download failed, retrying (4/5): Get "https://crpi-vofi3w62lkohhxsp.cn-s>
r msg="Download failed after 5 attempts: dial tcp: lookup crpi-vofi3w62lkohhxsp>
msg="Download failed, retrying (1/5): read tcp 172.19.70.232:45736->183.131.23>
msg="Download failed, retrying (2/5): read tcp 172.19.70.232:36172->183.131.23>
msg="Download failed, retrying (3/5): read tcp 172.19.70.232:55288->183.131.23>

[myhloli]

$ skopeo list-tags docker://crpi-vofi3w62lkohhxsp.cn-shanghai.personal.cr.aliyuncs.com/opendata
lab-mineru/ppu
{
    "Repository": "crpi-vofi3w62lkohhxsp.cn-shanghai.personal.cr.aliyuncs.com/opendatalab-mineru/ppu",
    "Tags": [
        "ppu-pytorch2.6.0-ubuntu24.04-cuda12.6-vllm0.8.5-py312"
    ]
}

我这边测试正常的，先检查下机器是否可以访问公网呢

[licyclone]

谢谢！内网网络问题，通过云上可以打上镜像了。另外有没有ppu的docker-compose的启动yml文件案例可以参考？

[licyclone]

services:
mineru-api:
image: mineru:ppu-vllm-latest
container_name: mineru-api
restart: always
ports:
- 8009:8000
environment:
MINERU_MODEL_SOURCE: local
#CUDA_VISIBLE_DEVICES: 2
entrypoint: mineru-api
command:
--host 0.0.0.0
--port 8000
--data-parallel-size 1
--gpu-memory-utilization 0.25 # If running on a single GPU and encountering VRAM shortage, reduce the KV cache size by this parameter, if VRAM issues persist, try lowering it further to 0.4 or below.
shm_size: 500g
ulimits:
memlock: -1
stack: 67108864
devices:
- /dev/alixpu:/dev/alixpu
- /dev/alixpu_ctl:/dev/alixpu_ctl
volumes:
- /mnt:/mnt
- /datapool:/datapool
- /var/run/docker.sock:/var
/run/docker.sock
ipc: host
privileged: true
deploy:
resources:
reservations:
devices:
- driver: nvidia
device_ids: [ "2" ]
capabilities: [ gpu ]

[whwususu]

昇腾910B环境的docker。同一个模型部署两套，一份在notebook,一份在ai服务，识别的文档一样。在notebook中用python命令的方式调用ocr识别正常，但是在ai服务器中通过python程序调用时，返回值部分段落是感叹号和少量乱码，什么原因？命令是mineru -p 输入.pdf -o 输出目录 这种模式。方式是vlm-http-client。 英伟达下正常，昇腾下，notebook下命令调用正常，python程序调用部分段落异常，有感叹号和其它乱码。
python调用代码：
file_name = str(Path(path).stem)
pdf_bytes = read_fn(path)
new_pdf_bytes = convert_pdf_bytes_to_bytes_by_pypdfium2(pdf_bytes, start_page_id, end_page_id)
backend = backend[4:]
middle_json, infer_result = vlm_doc_analyze(
new_pdf_bytes,
image_writer=None,
backend=backend,
server_url=server_url,
callback_func=bound_callback
)

notebook调用命令：
python demo.py -i pdfs -o out4

mineru版本为：2.6.5 启用的是vlm服务，镜像用的是官方镜像

[Chinaniu]

a3 npu模式部署minerU 解析文档 docker日志报 NNAL/ATB 缺失是什么情况 这个是我按照教程https://opendatalab.github.io/MinerU/zh/usage/acceleration_cards/Ascend/来更新的dockerfile，部分增加是我要弄缓存在改成 下面这种 FROM quay.m.daocloud.io/ascend/vllm-ascend:v0.11.0-a3

ENV PIP_INDEX_URL=https://mirrors.aliyun.com/pypi/simple
ENV HF_ENDPOINT=https://hf-mirror.com
ENV HF_HOME=/root/.cache/huggingface
ENV TRANSFORMERS_CACHE=/root/.cache/huggingface
ENV HF_HUB_ENABLE_HF_TRANSFER=1
ENV MINERU_MODEL_SOURCE=modelscope

RUN apt-get update &&
apt-get install -y
fonts-noto-core
fonts-noto-cjk
fontconfig
libgl1
libglib2.0-0 &&
fc-cache -fv &&
apt-get clean &&
rm -rf /var/lib/apt/lists/*

RUN python3 -m pip install -U pip -i $PIP_INDEX_URL &&
python3 -m pip install 'mineru[core]>=2.6.5' numpy==1.26.4 opencv-python==4.11.0.86 -i $PIP_INDEX_URL &&
python3 -m pip cache purge

ENTRYPOINT ["/bin/bash", "-c", "export MINERU_MODEL_SOURCE=local && exec "$@"", "--"] 运行命令是 下面这个我想用npu的形式解密以 gradio的方式运行 docker run -d --name mineru --privileged --ipc=host --restart unless-stopped -p 7860:7860 --device=/dev/davinci6 --device=/dev/davinci7 --device=/dev/davinci_manager --device=/dev/devmm_svm --device=/dev/hisi_hdc -v /usr/local/Ascend/driver:/usr/local/Ascend/driver -v /var/log/npu/:/usr/slog -v /data:/data -v /output:/output -v /newcache/modelscope:/root/.cache/modelscope -v /newcache/hf:/root/.cache/huggingface -v /newcache/mineru.json:/etc/trae-mineru.json:ro -e ASCEND_RT_VISIBLE_DEVICES=6,7 -e MINERU_MODEL_SOURCE=local -e MINERU_DEVICE_MODE=npu -e MINERU_TOOLS_CONFIG_JSON=/etc/trae-mineru.json -e LD_LIBRARY_PATH=/usr/local/Ascend/driver/lib64:$LD_LIBRARY_PATH -e TORCH_DEVICE_BACKEND_AUTOLOAD=0 mineru:npu-vllm-a3 mineru-gradio --server-name 0.0.0.0 --server-port 7860 啥原因导致的呢，是我改的dockerfile有问题吗

[vvu94271-png]

各位开发者好，我想请教一下，我现在使用的显卡是寒武纪 MLU370-X8 双卡，在官方最新的支持上，使用的是MLU590-M9D的显卡，现在的版本是否支持mlu370，如果不支持下一步有没有适配mlu370的计划
[2026-03-23 15:31:04.121247][CNNL][ERROR][32][Card:0]: [cnnlCreate] An internal error occured. When convert isa version to device failed. return unknown device. isa version = 372

[1556900941lizerui]

请教一下目前在310P上部署了mineru，貌似不支持数据并行，请求第二次的时候会报错，然后崩溃。现在我自己写了一个fastapi的服务，代码如下：但是速度特别特别慢，处理一页的图像都是二三十秒，是哪里有问题吗：
"""
FastAPI 服务 - 使用 MinerU 原生 http-client 后端异步调用 VLM 服务
基于 demo.py 逻辑，支持可调节信号量控制并发
"""

import os
import sys
import re

import time
import asyncio
import io
from pathlib import Path
from typing import List, Optional
from contextlib import asynccontextmanager
from datetime import datetime

from fastapi import FastAPI, UploadFile, File, Form, HTTPException
from fastapi.responses import JSONResponse
from loguru import logger
from PIL import Image

============ 日志配置 ============

LOG_DIR = Path(os.getenv("MINERU_LOG_DIR", "/mnt/share/lzr/code/CV/MinerU/claude_test/logs"))
LOG_DIR.mkdir(parents=True, exist_ok=True)
LOG_FILE = LOG_DIR / "vlm_api_server.log"

logger.remove()
logger.add(
sys.stderr,
level="INFO",
format="{time:YYYY-MM-DD HH:mm:ss} | {level: <8} | {name}:{function}:{line} - {message}"
)
logger.add(
str(LOG_FILE),
rotation="20 MB",
retention=5,
encoding="utf-8",
level="DEBUG",
format="{time:YYYY-MM-DD HH:mm:ss.SSS} | {level: <8} | {name}:{function}:{line} - {message}"
)

============ 导入 MinerU 组件 ============

from mineru.cli.common import convert_pdf_bytes_to_bytes_by_pypdfium2
from mineru.backend.vlm.vlm_analyze import aio_doc_analyze
from mineru.backend.vlm.vlm_middle_json_mkcontent import union_make as vlm_union_make
from mineru.utils.enum_class import MakeMode
from mineru.utils.pdf_image_tools import images_bytes_to_pdf_bytes
from mineru.utils.guess_suffix_or_lang import guess_suffix_by_bytes

============ 配置区域 ============

VLM_SERVER_URL = os.getenv("MINERU_VLM_SERVER_URL", "http://127.0.0.1:30000")
MAX_CONCURRENT_REQUESTS = int(os.getenv("MINERU_MAX_CONCURRENT", "12"))

文件类型定义

PDF_SUFFIXES = ["pdf"]
IMAGE_SUFFIXES = ["png", "jpeg", "jp2", "webp", "gif", "bmp", "jpg", "tiff"]

全局信号量

_request_semaphore: Optional[asyncio.Semaphore] = None

def convert_to_pdf_bytes(file_bytes: bytes, suffix: str) -> bytes:
"""将文件转换为 PDF bytes（图片会被转换）"""
if suffix in IMAGE_SUFFIXES:
try:
# 使用 PIL 转换图片为 PDF
image = Image.open(io.BytesIO(file_bytes))
# 根据 EXIF 信息自动转正
from PIL import ImageOps
image = ImageOps.exif_transpose(image) or image
# 转换为 RGB（PDF 不支持透明度）
if image.mode in ('RGBA', 'LA', 'P'):
background = Image.new('RGB', image.size, (255, 255, 255))
if image.mode == 'P':
image = image.convert('RGBA')
if image.mode in ('RGBA', 'LA'):
background.paste(image, mask=image.split()[-1] if image.mode in ('RGBA', 'LA') else None)
image = background
elif image.mode != 'RGB':
image = image.convert('RGB')

        # 保存为 PDF
        pdf_buffer = io.BytesIO()
        image.save(pdf_buffer, format='PDF', resolution=200.0)
        pdf_bytes = pdf_buffer.getvalue()
        pdf_buffer.close()
        return pdf_bytes
    except Exception as e:
        logger.error(f"图片转 PDF 失败: {e}")
        raise ValueError(f"图片转换失败: {e}")
elif suffix in PDF_SUFFIXES:
    return file_bytes
else:
    raise ValueError(f"不支持的文件类型: {suffix}")

def get_timestamp() -> str:
return datetime.now().strftime("%Y-%m-%d %H:%M:%S.%f")[:-3]

def format_duration(seconds: float) -> str:
if seconds < 1:
return f"{seconds * 1000:.2f}ms"
elif seconds < 60:
return f"{seconds:.2f}s"
else:
minutes = int(seconds // 60)
secs = seconds % 60
return f"{minutes}m {secs:.2f}s"

@asynccontextmanager
async def lifespan(app: FastAPI):
"""应用生命周期管理"""
global _request_semaphore
start_time = time.time()

logger.info("=" * 60)
logger.info("【服务启动】")
logger.info(f"  启动时间: {get_timestamp()}")
logger.info(f"  VLM 服务地址: {VLM_SERVER_URL}")
logger.info(f"  最大并发数: {MAX_CONCURRENT_REQUESTS}")
logger.info(f"  日志文件: {LOG_FILE}")
logger.info("=" * 60)

_request_semaphore = asyncio.Semaphore(MAX_CONCURRENT_REQUESTS)

yield

end_time = time.time()
logger.info("=" * 60)
logger.info("【服务关闭】")
logger.info(f"  关闭时间: {get_timestamp()}")
logger.info(f"  运行时长: {format_duration(end_time - start_time)}")
logger.info("=" * 60)

app = FastAPI(
title="MinerU VLM 解析服务",
description="使用 MinerU http-client 后端异步调用 VLM 服务",
version="2.0.0",
lifespan=lifespan
)

async def process_single_file(
file_bytes: bytes,
filename: str,
file_suffix: str,
start_page_id: int = 0,
end_page_id: Optional[int] = None
) -> dict:
"""
使用 MinerU 原生 http-client 后端处理单个文件
支持 PDF 和图片格式
返回按页分隔的 Markdown，去除图片
"""
global _request_semaphore

total_start = time.time()
logger.info(f"【文件处理开始】文件名: {filename} | 类型: {file_suffix} | 大小: {len(file_bytes)} bytes | 时间: {get_timestamp()}")

async with _request_semaphore:
    semaphore_start = time.time()
    logger.info(f"[信号量获取] 文件: {filename}")

    try:
        # 步骤1: 转换为 PDF bytes（图片会被转换）
        step1_start = time.time()
        pdf_bytes = convert_to_pdf_bytes(file_bytes, file_suffix)
        logger.info(f"[步骤1:格式转换] 文件: {filename} | {file_suffix} -> pdf | 输出大小: {len(pdf_bytes)} bytes")

        # 步骤2: 页面裁剪
        step2_start = time.time()
        pdf_bytes = convert_pdf_bytes_to_bytes_by_pypdfium2(pdf_bytes, start_page_id, end_page_id)
        step2_end = time.time()
        logger.info(f"[步骤2:页面裁剪] 文件: {filename} | 耗时: {format_duration(step2_end - step2_start)}")

        # 步骤3: 调用 MinerU 的 http-client 后端
        step3_start = time.time()
        logger.info(f"[步骤3:VLM解析] 文件: {filename} | 开始调用 http-client 后端")
        logger.info(f"[步骤3:VLM解析] 服务地址: {VLM_SERVER_URL}")

        try:
            # 传入 None 作为 image_writer，不写入图片
            middle_json, infer_result = await aio_doc_analyze(
                pdf_bytes=pdf_bytes,
                image_writer=None,  # 不写入图片
                backend="http-client",
                server_url=VLM_SERVER_URL,
                max_concurrency=MAX_CONCURRENT_REQUESTS,
                http_timeout=600,
            )
        except Exception as conn_err:
            logger.error(f"[步骤3:VLM解析] 连接 VLM 服务失败: {VLM_SERVER_URL}")
            logger.error(f"[步骤3:VLM解析] 错误: {str(conn_err)}")
            raise HTTPException(
                status_code=503,
                detail=f"无法连接到 VLM 服务 ({VLM_SERVER_URL})，请确认服务已启动"
            )

        step3_end = time.time()
        vlm_duration = step3_end - step3_start
        logger.info(f"[步骤3:VLM解析] 文件: {filename} | 耗时: {format_duration(vlm_duration)}")

        # 步骤4: 按页生成 Markdown 并合并
        step4_start = time.time()

        # 检查 middle_json 结构
        logger.debug(f"[调试] middle_json 类型: {type(middle_json)}")
        if isinstance(middle_json, str):
            logger.error(f"[错误] middle_json 是字符串而非字典: {middle_json[:200]}")
            raise ValueError("解析结果格式错误: middle_json 为字符串")

        pdf_info = middle_json.get("pdf_info", [])
        logger.debug(f"[调试] pdf_info 类型: {type(pdf_info)}, 页数: {len(pdf_info) if isinstance(pdf_info, list) else 'N/A'}")

        # 如果 pdf_info 不是列表，报错
        if isinstance(pdf_info, str):
            logger.error(f"[错误] pdf_info 是字符串: {pdf_info}")
            raise ValueError(f"pdf_info 格式错误，期望列表但得到字符串: {pdf_info}")
        if not isinstance(pdf_info, list):
            logger.error(f"[错误] pdf_info 类型错误: {type(pdf_info)}")
            raise ValueError(f"pdf_info 格式错误，期望列表但得到 {type(pdf_info)}")

        # 按页生成 Markdown
        page_markdowns = []
        for page_idx, page_info in enumerate(pdf_info):
            # 检查 page_info 类型
            if isinstance(page_info, str):
                logger.error(f"[错误] 页面 {page_idx} 是字符串: {page_info}")
                continue
            if not isinstance(page_info, dict):
                logger.error(f"[错误] 页面 {page_idx} 类型错误: {type(page_info)}")
                continue

            # 创建单页 pdf_info (传入列表而不是字典)
            single_page_info = [page_info]

            # 生成单页 Markdown（传入空字符串作为 image_dir，不处理图片路径）
            page_md = vlm_union_make(single_page_info, MakeMode.MM_MD, "")
            page_markdowns.append(page_md)

            logger.debug(f"[页面 {page_idx}] Markdown 长度: {len(page_md)} chars")

        # 用 dzint_page_sep 连接各页
        combined_md = "dzint_page_sep".join(page_markdowns)

        # 去除图片
        pattern = r'(\n+)?\s*<div\s+style="text-align:\s+center;"><file\s+src="[^"]*"\s+alt="Image"[^>]*>\s*</div>'
        final_content = re.sub(pattern, '', combined_md)

        step4_end = time.time()
        logger.info(f"[步骤4:生成Markdown] 文件: {filename} | 页数: {len(page_markdowns)} | 耗时: {format_duration(step4_end - step4_start)}")

        # 计算总耗时
        semaphore_end = time.time()
        semaphore_duration = semaphore_end - semaphore_start
        total_duration = semaphore_end - total_start

        logger.info(f"[信号量释放] 文件: {filename} | 信号量内耗时: {format_duration(semaphore_duration)}")
        logger.info(f"【文件处理完成】文件名: {filename} | 总耗时: {format_duration(total_duration)}")

        return {
            "markdown_text": final_content,
            "ocr_list": [],
            "page_count": len(page_markdowns),
            "processing_time": {
                "total_seconds": round(total_duration, 3),
                "total_formatted": format_duration(total_duration),
                "vlm_parse_seconds": round(vlm_duration, 3),
                "started_at": get_timestamp(),
                "completed_at": get_timestamp()
            }
        }

    except Exception as e:
        semaphore_end = time.time()
        total_duration = semaphore_end - total_start
        logger.exception(f"【文件处理失败】文件名: {filename} | 错误: {str(e)} | 耗时: {format_duration(total_duration)}")
        raise

@app.post("/parse")
async def parse_document(
file: UploadFile = File(..., description="PDF 或图片文件"),
language: str = Form("ch", description="语言"),
start_page_id: int = Form(0, description="起始页码"),
end_page_id: int = Form(99999, description="结束页码")
):
"""
解析单个文档（支持 PDF 和图片）
"""
request_start = time.time()
request_id = f"{int(request_start * 1000)}"

logger.info(f"【API请求开始】请求ID: {request_id} | 接口: /parse | 时间: {get_timestamp()}")
logger.info(f"[请求参数] 文件名: {file.filename} | 语言: {language}")

# 检查文件类型
allowed_extensions = ['.pdf', '.png', '.jpg', '.jpeg', '.webp', '.bmp', '.tiff', '.gif']
file_ext = Path(file.filename).suffix.lower()

if file_ext not in allowed_extensions:
    raise HTTPException(status_code=400, detail=f"不支持的文件类型: {file_ext}")

# 读取文件
content = await file.read()
if not content:
    raise HTTPException(status_code=400, detail="文件内容为空")

# 检测文件实际类型
file_suffix = guess_suffix_by_bytes(content)
logger.info(f"[文件检测] 文件名: {file.filename} | 检测类型: {file_suffix}")

try:
    # 处理文件
    result = await process_single_file(
        file_bytes=content,
        filename=file.filename,
        file_suffix=file_suffix,
        start_page_id=start_page_id,
        end_page_id=end_page_id if end_page_id < 99999 else None
    )

    request_end = time.time()
    logger.info(f"【API请求完成】请求ID: {request_id} | 总耗时: {format_duration(request_end - request_start)}")

    return JSONResponse({
        "markdown_text": result["markdown_text"],
        "ocr_list": result["ocr_list"],
        "page_count": result["page_count"],
        "processing_time": result["processing_time"],
        "status": "success"
    })

except Exception as e:
    logger.exception(f"处理失败: {file.filename}")
    raise HTTPException(status_code=500, detail=f"处理失败: {str(e)}")

@app.post("/parse/batch")
async def parse_documents_batch(
files: List[UploadFile] = File(..., description="多个文件（支持 PDF 和图片）"),
language: str = Form("ch")
):
"""批量解析（支持 PDF 和图片混合）"""
request_start = time.time()
request_id = f"batch_{int(request_start * 1000)}"

logger.info(f"【批量请求开始】请求ID: {request_id} | 文件数: {len(files)}")

if not files or len(files) > 20:
    raise HTTPException(status_code=400, detail="文件数必须在 1-20 之间")

# 读取所有文件并检测类型
file_list = []  # (filename, file_bytes, file_suffix)
for file in files:
    content = await file.read()
    if content:
        file_suffix = guess_suffix_by_bytes(content)
        file_list.append((file.filename, content, file_suffix))
        logger.info(f"[批量文件检测] {file.filename} -> {file_suffix}")

# 创建任务列表
tasks = []
for filename, content, file_suffix in file_list:
    task = process_single_file(
        file_bytes=content,
        filename=filename,
        file_suffix=file_suffix
    )
    tasks.append((filename, task))

# 并发执行
results = await asyncio.gather(*[t[1] for t in tasks], return_exceptions=True)

# 处理结果
response_data = []
success_count = 0
error_count = 0

for (filename, _), result in zip(tasks, results):
    if isinstance(result, Exception):
        error_count += 1
        response_data.append({
            "filename": filename,
            "status": "error",
            "error": str(result)
        })
    else:
        success_count += 1
        response_data.append({
            "filename": filename,
            "markdown_text": result["markdown_text"],
            "ocr_list": result["ocr_list"],
            "page_count": result["page_count"],
            "processing_time": result["processing_time"],
            "status": "success"
        })

request_end = time.time()
logger.info(f"【批量请求完成】请求ID: {request_id} | 成功: {success_count} | 失败: {error_count} | 总耗时: {format_duration(request_end - request_start)}")

return JSONResponse({
    "request_id": request_id,
    "total": len(files),
    "success": success_count,
    "failed": error_count,
    "total_time": format_duration(request_end - request_start),
    "results": response_data
})

@app.get("/health")
async def health_check():
return {"status": "healthy", "vlm_server": VLM_SERVER_URL, "max_concurrent": MAX_CONCURRENT_REQUESTS}

if name == "main":
import uvicorn
host = os.getenv("MINERU_API_HOST", "0.0.0.0")
port = int(os.getenv("MINERU_API_PORT", "8000"))
logger.info(f"启动服务: http://{host}:{port}")
uvicorn.run(app, host=host, port=port)